Recognition Is Solved. Discrimination Gates Autonomous Killing.

The bottleneck in autonomous weapons is not recognition, it is discrimination. A drone already classifies a vehicle or person at high accuracy. The unsolved problem is telling a soldier from a civilian from the air, and that single judgment separates AI-assisted targeting from a machine that kills on its own authority.

Recognition is a perception problem, and perception is what the last decade of deep learning has been quietly winning. Discrimination is a judgment problem, and judgment from the air needs context the model is never allowed to learn: every training set that distinguishes "civilian" is built from the same wars the model is meant to end. The US formally declines to build fully autonomous lethal weapons, on ethical grounds. The engineering is being done in combat regardless.

Alexander Palamarchuk, an R&D engineer in Ukraine's Azov Brigade working about 18 km from the front in the Pokrovsk region, puts the discrimination gap at roughly two years. That clock is shorter than any policy review, which means the ethical line is doing no work where the technology is actually being made. NATO's software advantage, Palamarchuk argues, lets Western countries win the autonomy race before competitors. That is a constructive frame, but it concedes the race itself. The mechanism is straightforward: the unit that closes discrimination first sets the standard the world inherits. Two years is a short window to decide who writes that standard, and whether civilians appear in it at all.